Read Devanagari from a photo or scan: the शिरोरेखा headline bar joining letters across a word, and conjunct clusters where consonants combine into a single glyph, are what this pack is trained on. It covers Hindi, Marathi, Nepali, Sanskrit, Bhojpuri, Maithili and Konkani, plus English on the same page.
How it works
- Drop an image or PDF. Text detection runs immediately.
- Pick your language pack (downloaded once, 8–17 MB) and press Read text.
- Copy the lines, or download .txt / .json. PDFs are read cover to cover.
Questions
- Does it handle conjuncts and matras correctly?
- That is what the Devanagari pack exists for — a general model trained on Latin text reads the headline bar as an underline and splits conjuncts into the wrong consonants. Output is Unicode Devanagari you can paste anywhere.
- Does my file get uploaded?
- No. The models and all processing run inside your browser on your own device. You can verify this yourself: open your browser’s DevTools, watch the Network tab, and process a file: there are no uploads. After the first visit most tools also work fully offline.
- Which languages are supported?
- 106 languages across Latin, Cyrillic, Arabic, Devanagari, CJK, Greek, Thai, Tamil and Telugu scripts. Bengali is the one major gap, since no v5 model exists for it yet, and the picker says so rather than guessing.
- How accurate is it?
- On clean document text, the measured character error rate is under 2%, and every line carries a confidence score so you can see which ones to check. Handwriting and low-resolution photos will read worse.
- Does it read scanned PDFs?
- Yes. Digital PDFs are read from their own text layer, with exact characters at exact positions. Scanned pages go through detection and recognition, and mixed documents use both.